user-manual: Add section "Why bisecting merge commits can be harder ..."

This commit adds a discussion of the challenge of bisecting
merge commits to the user manual.  The original author is
Junio C Hamano <gitster@pobox.com>, who posted the text to
the mailing list <http://marc.info/?l=git&m=119403257315527&w=2>.
His email was adapted for the manual.

The discussion is added to "Rewriting history and maintainig
patch series".  The text added requires good understanding of
merging and rebasing.  Therefore it should not be placed too
early in the manual.  Right after the section on "Problems with
rewriting history", the discussion of bisect gives another reason
for linearizing as much of the history as possible.

The text includes suggestions and fixes by
Ralf Wildenhues <Ralf.Wildenhues@gmx.de> and
Benoit Sigoure <tsuna@lrde.epita.fr>.

Signed-off-by: Steffen Prohaska <prohaska@zib.de>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>

Documentation/user-manual.txt

@@ -2554,6 +2554,72 @@ branches into their own work.
 For true distributed development that supports proper merging,
 published branches should never be rewritten.
 
+[[bisect-merges]]
+Why bisecting merge commits can be harder than bisecting linear history
+-----------------------------------------------------------------------
+
+The gitlink:git-bisect[1] command correctly handles history that
+includes merge commits.  However, when the commit that it finds is a
+merge commit, the user may need to work harder than usual to figure out
+why that commit introduced a problem.
+
+Imagine this history:
+
+................................................
+      ---Z---o---X---...---o---A---C---D
+          \                       /
+           o---o---Y---...---o---B
+................................................
+
+Suppose that on the upper line of development, the meaning of one
+of the functions that exists at Z is changed at commit X.  The
+commits from Z leading to A change both the function's
+implementation and all calling sites that exist at Z, as well
+as new calling sites they add, to be consistent.  There is no
+bug at A.
+
+Suppose that in the meantime on the lower line of development somebody
+adds a new calling site for that function at commit Y.  The
+commits from Z leading to B all assume the old semantics of that
+function and the callers and the callee are consistent with each
+other.  There is no bug at B, either.
+
+Suppose further that the two development lines merge cleanly at C,
+so no conflict resolution is required.
+
+Nevertheless, the code at C is broken, because the callers added
+on the lower line of development have not been converted to the new
+semantics introduced on the upper line of development.  So if all
+you know is that D is bad, that Z is good, and that
+gitlink:git-bisect[1] identifies C as the culprit, how will you
+figure out that the problem is due to this change in semantics?
+
+When the result of a git-bisect is a non-merge commit, you should
+normally be able to discover the problem by examining just that commit.
+Developers can make this easy by breaking their changes into small
+self-contained commits.  That won't help in the case above, however,
+because the problem isn't obvious from examination of any single
+commit; instead, a global view of the development is required.  To
+make matters worse, the change in semantics in the problematic
+function may be just one small part of the changes in the upper
+line of development.
+
+On the other hand, if instead of merging at C you had rebased the
+history between Z to B on top of A, you would have gotten this
+linear history:
+
+................................................................
+    ---Z---o---X--...---o---A---o---o---Y*--...---o---B*--D*
+................................................................
+
+Bisecting between Z and D* would hit a single culprit commit Y*,
+and understanding why Y* was broken would probably be easier.
+
+Partly for this reason, many experienced git users, even when
+working on an otherwise merge-heavy project, keep the history
+linear by rebasing against the latest upstream version before
+publishing.
+
 [[advanced-branch-management]]
 Advanced branch management
 ==========================